Wencan Jiang

I am a researcher at Zhejiang University. My work focuses on multimodal agents, long-horizon game control, and efficient reasoning for interactive environments.

My recent projects study how agents can act over long horizons under limited token and latency budgets. Instead of asking a large model to deliberate at every step, I am interested in systems that allocate reasoning to the moments where it matters: planning bottlenecks, state changes, failures, and recovery.

Research Interests

• Multimodal agents and long-horizon decision making
• Game agents and open-world interaction
• Efficient reasoning, memory, and agent control
• Vision-language models for interactive environments

Selected Projects

SPIKE: An Adaptive Dual Controller Framework for Cost-Efficient Long-Horizon Game Agents
SPIKE uses event-triggered dual control and hierarchical memory to improve the success-cost trade-off of long-horizon game agents. The project page is available here.

IAMFlow: Advancing Narrative Long Video Generation via Training-Free Identity-Aware Memory
IAMFlow studies identity-aware memory for long narrative video generation and persistent entity tracking across evolving prompts.

News

• May 2026: Released the SPIKE project page and paper materials.
• May 2026: Released IAMFlow and NarraStream-Bench for narrative long video generation.